Awesome

Scheduling APIs

This document outlines the motivation for working on various scheduling APIs<sup>1</sup>, discusses some of the problems that apps and userspace schedulers face when writing scheduling code, and links to various proposals we are working on in this space.

(<sup>1</sup>The scope of this work was previously restricted to main-thread scheduling, and while main-thread scheduling remains the primary focus, the repository and some accompanying text has been renamed to "scheduling-apis" to reflect the inclusion of APIs like scheduler.postTask() on workers.)

Motivation: Main-thread Contention

Applications may experience main-thread contention at various points in their execution, e.g. during page load or as a result of user interaction. This contention can negatively affect user experience in terms of responsiveness and latency. For example, a busy main thread can prevent the UA from servicing input, leading to poor responsiveness. Similarly, tasks (e.g. fetch completions, rendering, etc.) can experience large queuing durations during times of contention, which increases task latency and can result in degraded quality of experience.

Consider a "search-as-you-type" application. This app needs to be responsive to user input, i.e. users typing in the search-box. At the same time, any animations on the page must be rendered smoothly, and the work for fetching and preparing search results and updating the page must also progress quickly. There are a lot of different deadlines to meet for the app developer. It is easy for any long running script work to hold up the main thread and cause responsiveness issues for typing, rendering animations, or updating search results.

Another example pinch-zooming in a map application. The app needs to continuously respond to the input, update the rendering, and potentially fetch new content to be displayed. Similar to the search-as-you-type example, long running script work could block other tasks, making the application feel laggy.

Current Solutions, Their Limitations, and APIs to Fill the Gaps

Dealing with contention is largely a scheduling problem: to the degree that work can be reordered in an more optimal way, scheduling can have a positive impact. What makes this problem more pronounced on the web is that tasks run to completion—the UA cannot preempt a task to run high priority work like processing user input. This problem is generally tackled in userspace by systematically chunking and scheduling main-thread work. Since long tasks and responsiveness are at odds, breaking up long tasks can help keep an app responsive when also yielding to the browser's event loop.

Userspace schedulers have evolved to manage these chunks of work—prioritizing and executing work async at an appropriate time relative to current situation of user and browser. And while userspace schedulers have been effective in improving responsiveness, there are several problems they still face:

Coordination between (cooperating) actors: Most userspace schedulers have a notion of priority that allows tasks to be ordered in a way that improves user experience. But this is limited since userspace schedulers do not control all tasks on the page.

Apps can consist of 1P, 1P library, 3P, and (one or more) framework script each of which competes for the main thread. At the same time, the browser also has tasks to run on the main thread, such as fetch() and IDB tasks and garbage collection.

Having a shared notion of priority can help the browser make better scheduling decisions, which in turn can help improve user experience. We propose adding a prioritized task scheduling API to address this problem.
A disparate set of scheduling APIs: Despite the need to schedule chunks of script, the Platform lacks a unified API to do so. Developers can choose setTimeout, postMessage, requestAnimationFrame, or requestIdleCallback, when choosing to schedule tasks.

This disparate set of scheduling APIs makes it even more difficult for developers to write scheduling code and requires expert knowledge of the browser's event loop to do so. Creating a unified native scheduling API —scheduler.postTask() —will alleviate this.
Determining when to yield to the browser: yielding has overhead—the overhead of posting a task and context switching, the cost of regaining control, etc. This can lead to increased task latency.

Making intelligent decisions about when to yield is difficult with limited knowledge. Scheduling primitives can help userspace schedulers make better decisions, e.g. isInputPending() and isFramePending().
Regaining control after yielding: chunking work and yielding is necessary for improving responsiveness, but it comes at a cost: when yielding to the event loop, a task that yields has no way to continue without arbitrary work of the same priority running first, e.g. other script. This disincentivizes yielding from a script that requires low task latency. Providing a primitive like scheduler.yield() that is designed to take into account this async userspace task model can help, as the scheduler can prioritize these continuations more fairly.

Additional Scheduling Problems

The problem as described above only covers part of the scheduling problem space. Additionally, there are developer needs for things like detecting when a frame is pending, throttling the frame rate, and avoiding layout thrashing. Some of the other APIs we are considering in this space are noted here.

APIs and Status